We Teach a Computer to Learn ?

Traditionally, whenever we want to use a computer to solve a new task, we need to provide it with very explicit and precise description of what it is exactly that we want it to do. This requirement often is not too problematic – after all, there is a lot of tasks that even not too experienced programmer (let alone a team of experienced ones) can make computers perform. However, there still is a lot of scenarios in which just coming up with a description of the task at hand that is sufficiently explicit and precise for a computer might be the core of the problem. A canonical example here is email spam filtering. Namely, if you see an email message in your inbox, you can immediately tell if it is a spam or not. On the other hand, specifying – in a precise, succinct, and computer-readable way – what it means for an email message to be a spam might be rather challenging. Let us take a closer look at this problem. We can view each email message as an object that is described by a set of features: the From address, Reply-to address, Subject, presence of certain keywords in its body, etc. Now, the traditional approach to dealing with spam messages would be based on looking at the pool of the emails we got so far and providing our email system with a set of filtering rules that would help distinguish the emails in this pool that are spam from the ones that are not. For example, we could setup rules like: